rohan13's picture
Prompt changes for formatting response
4604a12
raw
history blame
20 kB
import os
import pickle
import langchain
import faiss
from langchain import HuggingFaceHub, PromptTemplate
from langchain.chains import ConversationalRetrievalChain, LLMChain
from langchain.chat_models import ChatOpenAI
from langchain.llms import OpenAI
from langchain.document_loaders import DirectoryLoader, TextLoader, UnstructuredHTMLLoader
from langchain.embeddings import OpenAIEmbeddings, HuggingFaceHubEmbeddings
from langchain.memory import ConversationBufferWindowMemory
from langchain.prompts.chat import (
ChatPromptTemplate,
HumanMessagePromptTemplate,
SystemMessagePromptTemplate,
StringPromptTemplate
)
from langchain.output_parsers import PydanticOutputParser
from langchain.tools.json.tool import JsonSpec
from typing import List, Union, Callable
from langchain.schema import AgentAction, AgentFinish
import re
from langchain.text_splitter import CharacterTextSplitter
from custom_faiss import MyFAISS
from langchain.cache import InMemoryCache
from langchain.chat_models import ChatGooglePalm
from langchain.document_loaders import JSONLoader
from langchain.agents import initialize_agent, Tool, AgentType
from langchain.agents import Tool, AgentExecutor, LLMSingleActionAgent, AgentOutputParser, BaseMultiActionAgent
from langchain.tools import StructuredTool
from langchain.chains import create_tagging_chain
from typing import List, Tuple, Any, Union
from langchain.schema import AgentAction, AgentFinish
from pydantic import BaseModel, Field
from typing import Optional
class ToolArgsSchema(BaseModel):
student_name: Optional[str] = Field(description="The name of the student")
question: str = Field(description="The question being asked")
question_type: str = Field(description="The type of question being asked")
interest: Optional[str] = Field(description="The interest of the student")
class Config:
schema_extra = {
"required": ["question", "question_type"]
}
langchain.llm_cache = InMemoryCache()
model_name = "GPT-4"
pickle_file = "_vs.pkl"
index_file = "_vs.index"
models_folder = "models/"
os.environ["LANGCHAIN_TRACING"] = "true"
discussions_file_path = "discussion_entries.json"
llm = OpenAI(model_name="gpt-3.5-turbo-16k", temperature=0, verbose=True)
embeddings = OpenAIEmbeddings(model='text-embedding-ada-002')
chat_history = []
memory = ConversationBufferWindowMemory(memory_key="chat_history", k=10)
vectorstore_index = None
agent_prompt = """
I am the LLM AI canvas discussion grading assistant.
I can answer two types of questions: grade-based questions and interest-based questions.
Grade-based questions are about the grades of a certain student or a group of students based on the rubric below for the canvas discussion on the topic 8 nouns. ALWAYS return total score when it is grading based question.
Interest-based questions are about the interests or skills of a certain student or a group of students based on their discussion posts.
You have access to the following tools:
{tools}
Use the following format:
Question: the input question you must answer
Thought: you should always think about type of question it is
Action: the action to take, should be one of [{tool_names}]
Action Input: the input to the action
Observation: the result of the action
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question
Begin!
Question: {input}
{agent_scratchpad}
"""
# Set up a prompt template
class CustomPromptTemplate(StringPromptTemplate):
# The template to use
template: str
############## NEW ######################
# The list of tools available
tools_getter: Callable
def format(self, **kwargs) -> str:
# Get the intermediate steps (AgentAction, Observation tuples)
# Format them in a particular way
intermediate_steps = kwargs.pop("intermediate_steps")
thoughts = ""
for action, observation in intermediate_steps:
thoughts += action.log
thoughts += f"\nObservation: {observation}\nThought: "
# Set the agent_scratchpad variable to that value
kwargs["agent_scratchpad"] = thoughts
############## NEW ######################
tools = self.tools_getter(kwargs["input"])
# Create a tools variable from the list of tools provided
kwargs["tools"] = "\n".join(
[f"{tool.name}: {tool.description}" for tool in tools]
)
# Create a list of tool names for the tools provided
kwargs["tool_names"] = ", ".join([tool.name for tool in tools])
return self.template.format(**kwargs)
class CustomOutputParser(AgentOutputParser):
def parse(self, llm_output: str) -> Union[AgentAction, AgentFinish]:
print("llm_output")
print(llm_output)
# Check if agent should finish
if "Final Answer:" in llm_output:
return AgentFinish(
# Return values is generally always a dictionary with a single `output` key
# It is not recommended to try anything else at the moment :)
return_values={"output": llm_output.split("Final Answer:")[-1].strip()},
log=llm_output,
)
# Parse out the action and action input
regex = r"Action\s*\d*\s*:(.*?)\nAction\s*\d*\s*Input\s*\d*\s*:[\s]*(.*)"
match = re.search(regex, llm_output, re.DOTALL)
if not match:
raise ValueError(f"Could not parse LLM output: `{llm_output}`")
action = match.group(1).strip()
action_input = match.group(2)
# Return the action and action input
return AgentAction(tool=action, tool_input=action_input.strip(" ").strip('"'), log=llm_output)
system_template = """
I am the LLM AI canvas discussion grading assistant.
I can answer two types of questions: grade-based questions and interest-based questions.
Grade-based questions are about the grades of a certain student or a group of students based on the rubric below for the canvas discussion on the topic 8 nouns.
Interest-based questions are about the interests or skills of a certain student or a group of students based on their discussion posts.
To grade student discussions, I will follow the rubric below.
Student Post
3 points: Post includes 8 nouns and text describing how these nouns relate to the student.
2 points: Student's post includes 8 nouns but does not offer how those nouns relate to the student.
1 point: Student's post has significant missing details.
0 points: The student does not provide an initial post, or otherwise does not follow assignment instructions.
Response to Others
3 points: Student responds to at least 3 other student discussion threads AND responds to questions asked of them. Student posts insightful comments that prompt on target discussion. These posts also avoid throw away comments such as I agree, Me too, Good idea.
2 points: Student was notably lacking in one criterion.
1 point: Student was notably lacking in two criteria.
0 points: The student does not interact in the threads of other students.
I will be able to identify each student by name, and I will be able to share their likings, interests, and other characteristics. I will also be able to filter out students based on their interests.
I will not deviate from the grading scheme. I will grade each discussion entry and reply carefully, and I will share the grades of all individuals by name on the basis of the rubric. I will ALWAYS return total score when it is grading based question.
The discussions and their replies are in following format:
Student Post: Student Name
Reply to: Another Student Discussion ID
Your answer to grade based questions should be in following format:
Student Post: X points
Response to Others: X points
Total: X points
Following are the relevant discussions to grade or answer the interest based questions
----------------
Discussions:
{context}"""
messages = [
SystemMessagePromptTemplate.from_template(system_template),
HumanMessagePromptTemplate.from_template("{question}"),
]
CHAT_PROMPT = ChatPromptTemplate.from_messages(messages)
def set_model_and_embeddings():
global chat_history
# set_model(model)
# set_embeddings(model)
chat_history = []
def set_embeddings(model):
global embeddings
if model == "GPT-3.5" or model == "GPT-4":
print("Loading OpenAI embeddings")
embeddings = OpenAIEmbeddings(model='text-embedding-ada-002')
elif model == "Flan UL2" or model == "Flan T5":
print("Loading Hugging Face embeddings")
embeddings = HuggingFaceHubEmbeddings(repo_id="sentence-transformers/all-MiniLM-L6-v2")
def get_search_index():
global vectorstore_index, model_name
if os.path.isfile(get_file_path(model_name, pickle_file)) and os.path.isfile(
get_file_path(model_name, index_file)) and os.path.getsize(get_file_path(model_name, pickle_file)) > 0:
# Load index from pickle file
with open(get_file_path(model_name, pickle_file), "rb") as f:
# search_index = Chroma(persist_directory=models_folder, embedding_function=embeddings)
search_index = pickle.load(f)
print("Loaded index")
else:
search_index = create_index(model_name)
print("Created index")
vectorstore_index = search_index
return search_index
def create_index(model):
source_chunks = create_chunk_documents()
search_index = search_index_from_docs(source_chunks)
# search_index.persist()
faiss.write_index(search_index.index, get_file_path(model, index_file))
# Save index to pickle file
with open(get_file_path(model, pickle_file), "wb") as f:
pickle.dump(search_index, f)
return search_index
def get_file_path(model, file):
# If model is GPT3.5 or GPT4 return models_folder + openai + file else return models_folder + hf + file
if model == "GPT-3.5" or model == "GPT-4":
return models_folder + "openai" + file
else:
return models_folder + "hf" + file
def search_index_from_docs(source_chunks):
# print("source chunks: " + str(len(source_chunks)))
# print("embeddings: " + str(embeddings))
search_index = MyFAISS.from_documents(source_chunks, embeddings)
return search_index
def get_html_files():
loader = DirectoryLoader('docs', glob="**/*.html", loader_cls=UnstructuredHTMLLoader, recursive=True)
document_list = loader.load()
for document in document_list:
document.metadata["name"] = document.metadata["source"].split("/")[-1].split(".")[0]
return document_list
def metadata_func(record: dict, metadata: dict) -> dict:
metadata["name"] = record.get("name")
return metadata
def get_json_file():
global discussions_file_path
loader = JSONLoader(
file_path=discussions_file_path,
jq_schema='.[]', metadata_func=metadata_func, content_key="message")
return loader.load()
def fetch_data_for_embeddings():
# document_list = get_text_files()
document_list = get_html_files()
# document_list = get_json_file()
print("document list: " + str(len(document_list)))
return document_list
def get_text_files():
loader = DirectoryLoader('docs', glob="**/*.txt", loader_cls=TextLoader, recursive=True)
document_list = loader.load()
return document_list
def create_chunk_documents():
sources = fetch_data_for_embeddings()
splitter = CharacterTextSplitter(separator=" ", chunk_size=800, chunk_overlap=0)
source_chunks = splitter.split_documents(sources)
print("chunks: " + str(len(source_chunks)))
return sources
def get_qa_chain(vectorstore_index, question, metadata):
global llm, model_name
print(llm)
filter_dict = {"name": metadata.student_name}
# embeddings_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.76)
# compression_retriever = ContextualCompressionRetriever(base_compressor=embeddings_filter, base_retriever=gpt_3_5_index.as_retriever())
retriever = get_retriever(filter_dict, vectorstore_index, metadata)
print(retriever.get_relevant_documents(question))
chain = ConversationalRetrievalChain.from_llm(llm, retriever, return_source_documents=True,
verbose=True, get_chat_history=get_chat_history,
combine_docs_chain_kwargs={"prompt": CHAT_PROMPT})
return chain
def get_retriever(filter_dict, vectorstore_index, metadata):
if metadata.question_type == "grade-based":
retriever = vectorstore_index.as_retriever(search_type='mmr',
search_kwargs={'lambda_mult': 1, 'fetch_k': 20, 'k': 10,
'filter': filter_dict})
else:
retriever = vectorstore_index.as_retriever(search_type='mmr',
search_kwargs={'lambda_mult': 1, 'fetch_k': 20, 'k': 10})
return retriever
def get_chat_history(inputs) -> str:
res = []
for human, ai in inputs:
res.append(f"Human:{human}\nAI:{ai}")
return "\n".join(res)
def generate_answer(question, metadata: ToolArgsSchema) -> str:
# print("filter: " + filter)
global chat_history, vectorstore_index
chain = get_qa_chain(vectorstore_index, question, metadata)
result = chain(
{"question": question, "chat_history": chat_history})
chat_history.extend([(question, result["answer"])])
sources = []
print(result)
for document in result['source_documents']:
source = document.metadata['source']
sources.append(source.split('/')[-1].split('.')[0])
print(sources)
source = ',\n'.join(set(sources))
# return result['answer'] + '\nSOURCES: ' + source
return result['answer']
def get_question_type(question):
parser = PydanticOutputParser(pydantic_object=ToolArgsSchema)
prompt_template = """I can answer two types of questions: grade-based questions and interest-based questions.
Grade-based questions are about the grades of a certain student or a group of students based on the rubric below for the canvas discussion on the topic 8 nouns.
Interest-based questions are about the interests or skills of a certain student or a group of students based on their discussion posts.
Question: {question}
Find following information about the question asked. Return Optional empty if the information is not available.:
Format instructions: {format_instructions}"""
llm = OpenAI(temperature=0)
prompt = PromptTemplate(template=prompt_template, input_variables=["question"], output_parser=parser, partial_variables={"format_instructions": parser.get_format_instructions()})
llm_chain = LLMChain(
llm=llm,
prompt=prompt,
)
output = llm_chain.run(question)
output = parser.parse(output)
output = generate_answer(question, output)
return output
# class FakeAgent(BaseMultiActionAgent):
# """Fake Custom Agent."""
#
# @property
# def input_keys(self):
# return ["input"]
#
# def plan(
# self, intermediate_steps: List[Tuple[AgentAction, str]], **kwargs: Any
# ) -> Union[List[AgentAction], AgentFinish]:
# print("input keys")
# print(self.input_keys)
# print("intermediate steps")
# print(intermediate_steps)
# print("kwargs")
# print(kwargs)
#
# """Given input, decided what to do.
#
# Args:
# intermediate_steps: Steps the LLM has taken to date,
# along with observations
# **kwargs: User inputs.
#
# Returns:
# Action specifying what tool to use.
# """
# if len(intermediate_steps) == 0:
# first_action = AgentAction(tool="question type", tool_input=kwargs["input"], log="")
# print("first action")
# print(first_action)
# second_action = AgentAction(tool="Grade",tool_input=kwargs["input"], log="")
# print("second action")
# print(second_action)
# return [
# first_action,
# second_action,
# ]
# else:
# return AgentFinish(return_values={"output": "bar"}, log="")
#
# async def aplan(
# self, intermediate_steps: List[Tuple[AgentAction, str]], **kwargs: Any
# ) -> Union[List[AgentAction], AgentFinish]:
# """Given input, decided what to do.
#
# Args:
# intermediate_steps: Steps the LLM has taken to date,
# along with observations
# **kwargs: User inputs.
#
# Returns:
# Action specifying what tool to use.
# """
# if len(intermediate_steps) == 0:
# return [
# AgentAction(tool="question type", tool_input=kwargs["input"], log=""),
# AgentAction(tool="Grade",
# tool_input={
# "student_name": kwargs["student_name"],
# "question": kwargs["question"],
# "question_type": kwargs["question_type"],
# "interest": kwargs["interest"]
# }, log=""),
# ]
# else:
# return AgentFinish(return_values={"output": "bar"}, log="")
#
#
# schema = {
# "properties": {
# "student_name" : {"type": "string", "description": "The name of the student"},
# "question": {"type": "string", "description": "The question being asked"},
# "question type" : {"type": "string",
# "enum": ["student grades", "student specific", "interest specific"],
# "description": "The type of question being asked"},
# "interest" : {"type": "string", "description": "The interest of the student"},
# },
# "required": ["question", "question type"]
# }
# def get_tagging_chain(question)-> str:
# global schema
# chain = create_tagging_chain(schema, llm)
# first_answer = chain.run(question)
# print("first answer:")
# print(first_answer)
# return first_answer
#
#
# def get_grading_agent():
#
# tools = [
# Tool(
# name="question type",
# func=get_tagging_chain,
# description="Useful when you need to understand the type of the input."
# ),
# StructuredTool(
# name="Grade",
# func=generate_answer,
# description="Useful when you need to answer questions about students, grades, interests, etc from the context of canvas discussion posts. If the question is student specific, student name is required.",
# args_schema=ToolArgsSchema
# )
# ]
# # agent = initialize_agent(tools, llm, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
#
# agent = FakeAgent(output_parser=CustomOutputParser())
# # prompt = CustomPromptTemplate(template=agent_prompt, tools=tools, input_variables=["input", "intermediate_steps"])
# # output_parser = CustomOutputParser()
# # tool_names = [tool.name for tool in tools]
# # llm_chain = LLMChain(llm=llm, prompt=prompt)
# # agent = LLMSingleActionAgent(
# # llm_chain=llm_chain,
# # output_parser=output_parser,
# # stop=["\nObservation:"],
# # allowed_tools=tool_names,
# # )
# agent_executor = AgentExecutor.from_agent_and_tools(
# agent=agent, tools=tools, verbose=True
# )
#
# # return initialize_agent(tools, llm, agent=AgentType.OPENAI_FUNCTIONS, verbose=True)
# return agent_executor
#
#
#
# def grade_answer(question) -> str:
# global chat_history, vectorstore_index
# agent = get_grading_agent()
# return agent.run(question)